Research map / 3D vision / Depth and stereo
EDM: Equirectangular Projection-Oriented Dense Kernelized Feature Matching
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EDM, a learning-based dense kernelized feature matcher for equirectangular (ERP) omnidirectional images using spherical positional embeddings, geodesic flow refinement on the unit sphere, and azimuth rotation augmentation.
Problem
Dense feature matching between omnidirectional (ERP) images is hard because ERP distortions, especially near poles, break perspective-based dense matchers like DKM and RoMa.
Approach
Reformulate dense kernelized matching on the unit sphere: embed features with 3D Cartesian spherical positional embeddings in the GP global matcher, refine warps by converting between Cartesian and ERP coordinates, and supervise with angular (cosine) regression loss, plus azimuth rotation augmentation.
Key results
- On Matterport3D, EDM reaches 45.15 AUC@5° versus 18.43 for DKM (+26.72).
- On Stanford2D3D (zero-shot from Matterport3D-trained model), EDM reaches 55.08 AUC@5°, +42.62 over the best prior method.
- Ablation: 3D Cartesian positional embedding plus bidirectional transformation plus rotational augmentation raises AUC@5° from 19.83 (DKM*) to 45.15.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EDM instead of DKM | Dense matching for omnidirectional images · Matterport3D test | AUC@5°: +26.72 | Helped |
| Use EDM instead of DKM | Dense matching for omnidirectional images · Stanford2D3D test | AUC@5°: +42.62 | Helped |
| Add bidirectional coordinate transformation for refinement (ablation) | EDM · Matterport3D test | AUC@5°: 19.83 -> 29.67 | Helped |
| Add rotational augmentation (ablation) | EDM with 2D linear positional embedding · Matterport3D test | AUC@5°: 29.67 -> 35.03 | Helped |
| Use 3D Cartesian linear positional embedding instead of 2D linear positional embedding (ablation) | EDM with bidirectional transformation and no rotational augmentation · Matterport3D test | AUC@5°: 29.67 -> 34.64 | Helped |
| Add rotational augmentation (ablation) | EDM with 3D linear positional embedding · Matterport3D test | AUC@5°: 34.64 -> 45.15 | Helped |
| Replace 3D linear positional embedding with 3D sinusoidal positional embedding (ablation) | EDM with bidirectional transformation and rotational augmentation · Matterport3D test | AUC@5°: 45.15 -> 42.39 | Hurt |
Surprising findings
- Pinhole-based DKM retrained with a spherical warp loss (DKM*) performs far worse than EDM, showing the architecture matters beyond the loss change.
- RoMa, a newer perspective dense matcher, underperforms DKM on Matterport3D ERP pose estimation (12.45 vs 18.43 AUC@5°).
Limitations
- Trained mainly on indoor datasets with vertically oriented cameras, so it is vulnerable to extreme rotations and outdoor scenes.
- Severely slanted image pairs from rotational or egocentric capture can cause failures.
- Evaluation on Stanford2D3D and qualitative sets relies on a model trained only on Matterport3D.
Experimental setup
Datasets: Matterport3D Stanford2D3D EgoNeRF OmniPhotos
Benchmarks: Matterport3D Stanford2D3D
Metrics: AUC@5° AUC@10° AUC@20°
Baselines: SPHORB SphereGlue DKM RoMa DKM*
Builds on: DKM RoMa LoFTR
Closest papers by meaning
- Spherical Pseudo-Cylindrical Representation for Omnidirectional Image Super-resolution (AAAI 2024 · similarity 0.84)
- Dense-SfM: Structure from Motion with Dense Consistent Matching (CVPR 2025 · similarity 0.82)
- HomoMatcher: Achieving Dense Feature Matching with Semi-Dense Efficiency by Homography Estimation (AAAI 2025 · similarity 0.82)
- Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed (CVPR 2024 · similarity 0.82)
- OmniGlue: Generalizable Feature Matching with Foundation Model Guidance (CVPR 2024 · similarity 0.82)
- UFM: A Simple Path towards Unified Dense Correspondence with Flow (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.